Pith. sign in

REVIEW 4 major objections 5 minor 66 references

The paper claims that a surrogate model for reactive CO2 flow in porous media can be trained on small 64×64 subdomains and then applied to the full 256×256 domain, cutting training memory and matching or beating reduced-order and full-domai

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 20:43 UTC pith:AWPS5NQD

load-bearing objection Useful applied benchmark for patch-trained fully convolutional surrogates in CO2 storage: memory savings and accuracy are real, but the 'unseen data' metrics are compromised by validation-set tuning and Table 1 contradicts the UNet++ conclusion. the 4 major comments →

arxiv 2602.22188 v2 pith:AWPS5NQD submitted 2026-02-25 cs.LG cs.AIphysics.flu-dyn

Surrogate models for Rock-Fluid Interaction: A Grid-Size-Invariant Approach

Nathalie C. Pinheiro , Donghu Guo , Hannah P. Menke , Aniket C. Joshi , Claire E. Heaney , Ahmed H. ElSheikh , Christopher C. Pain This is my paper
classification cs.LG cs.AIphysics.flu-dyn
keywords surrogate modelsgrid-size invariancefully convolutional networksUNet++reactive transportporous mediacarbon storagerollout training
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper aims to establish that a surrogate model for reactive CO2 flow in porous media can be trained on small 64×64 subdomains and then, because the network is fully convolutional, applied to the full 256×256 domain at inference. The grid-size-invariance property is presented as a way to slash GPU memory during training while keeping predictive accuracy, and the paper claims it outperforms the reduced-order models it compares against and generalizes better than a full-domain baseline on unseen simulations. The application is deliberately hard: injected CO2 dissolves the rock, so the solid geometry changes over time and cannot be used as a fixed mask to correct predictions. If the claim holds, the method makes large, finely resolved simulation datasets — including future 3D ones — tractable for multi-query tasks such as uncertainty quantification, and it identifies UNet++ with rollout training as the strongest configuration.

Core claim

The central discovery is that grid-size invariance — training a fully convolutional UNet or UNet++ on uniformly sampled 64×64 patches and inferring on the whole 256×256 field — produces surrogate predictions for CO2 concentration, porosity, and velocity whose Pearson correlation with ground truth stays above 0.75 over 100 autoregressive timesteps on validation simulations. The paper reports that these grid-size-invariant models outperform the reduced-order autoencoder-based surrogates on unseen data and also beat a UNet++ baseline trained on full-domain samples, interpreting the patch sampling as data augmentation that reduces overfitting. It further shows that rollout training, which penali

What carries the argument

The load-bearing mechanism is the fully convolutional network: with only local convolutional connections and no fully connected layers, a trained UNet or UNet++ can process inputs of arbitrary spatial size, so it can be trained on 64×64 subsamples and evaluated on a 256×256 grid with the same grid spacing. UNet is a convolutional encoder-decoder with skip connections; UNet++ adds nested convolutional blocks along the skip paths to refine multiscale features. Around this core, the paper adds rollout training (compute the loss over T=8 autoregressive steps, initialized by conventional single-step training), a boundary-condition loss term whose weight is gradually increased, and uniform spatial

Load-bearing premise

The load-bearing premise, stated in Section 2.2, is that the 64×64 training subsamples are representative of everything the model must later predict on the 256×256 domain — including the growth, branching, and reaction patterns that can extend beyond a single patch; if a feature only appears at scales larger than the patch, the grid-size-invariant model cannot recover it.

What would settle it

Train the same grid-size-invariant UNet/UNet++ on the same data with patch sizes 32×32, 64×64, and 128×128, then roll out 100 validation timesteps and compare PCC or SSIM; if performance drops sharply once the characteristic channel-bifurcation or dissolution-front length exceeds the patch size, the invariance is bounded by the subsample scale, contradicting the claim that the approach is a general route to larger domains.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • GPU-peak memory for training the predictor drops from about 1.8 GB (full-domain UNet) or 3.4 GB (full-domain UNet++) to 0.26–1.5 GB for the grid-size-invariant variants, so larger or 3D datasets become trainable on a single consumer GPU.
  • Inference on a 256×256 field takes under about 1.5 seconds for 97 autoregressive steps, versus roughly hours for the full-order simulation, enabling many-query workflows such as uncertainty quantification and optimisation.
  • Rollout training with T=8 improves long-horizon prediction: the best grid-size-invariant models keep PCC above 0.75 in all fields over 100 timesteps and reduce the error in predicted CO2 area and MSE relative to the no-rollout versions.
  • On validation data the grid-size-invariant models outperform the reduced-order models and the full-domain baseline, suggesting the subsampling acts as data augmentation against overfitting.
  • The framework is stated to be independent of this specific dataset and is being targeted at larger and 3D problems and other flow regimes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to sweep patch size (e.g., 32, 64, 128) on the same dataset; the patch size at which validation performance collapses would map the scale envelope of the invariance, which the paper does not quantify.
  • The method's memory savings apply during training only; inference uses the full domain, so the practical win is for training large models, not for cheap rollout on small machines.
  • Because subsampling is uniform and the paper notes samples must be representative, an adaptive sampler biased toward reaction fronts and channel bifurcations could improve accuracy, at the price of introducing sampling bias that would need to be controlled.
  • The stated plan to combine the surrogate with a PDE solver is natural: the PCC=0.75 threshold observed in the metrics can serve as a hand-back trigger, letting the surrogate fast-forward stable intervals and the solver guarantee accuracy when the trigger fires.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes and compares eight surrogate models for reactive CO2 transport in porous media. Four are reduced-order models (autoencoder or adversarial autoencoder compression followed by UNet or UNet++ latent-space prediction), and four are single fully convolutional UNet/UNet++ models trained on 64x64 subsamples and applied at 256x256, which the authors call grid-size-invariant. The central claims are that this grid-size-invariant approach reduces training memory, generalizes to unseen validation simulations, and outperforms the ROMs considered, and that UNet++ outperforms UNet. Results are reported as PCC, SSIM, MSE, and a CO2-area error metric over 8 validation simulations after 100-step autoregressive rollout.

Significance. If the grid-size-invariant claim holds, the paper would make a useful practical contribution: fully convolutional training on small patches is a cheap way to train image-to-image surrogates for large-domain inference, and the detailed memory and timing comparisons in Table 3 are valuable for practitioners. The comparison of UNet versus UNet++ and of traditional versus rollout training is also potentially informative. However, the current evidence is weakened by the reuse of the 8 validation simulations for model selection and by internal inconsistencies in the reported metrics, so the contributions are not yet demonstrated at the level claimed.

major comments (4)
  1. [§3.4, Table 1; Conclusions, item 2] The abstract and Conclusion 2 claim that UNet++ “resulted in better predictions” than UNet, but Table 1 directly contradicts this for the grid-size-invariant models. In Table 1, the best overall model is UNet with rollout (PCC for CO2 0.93, porosity 0.97, velocity-X 0.88; SSIM 0.81/0.91/0.94/0.95), while UNet++ with rollout has PCC 0.90/0.96/0.83/0.75 and SSIM 0.81/0.90/0.93/0.95. UNet++ is not uniformly better on any of the four fields. This inconsistency needs to be resolved either by changing the claim or by aggregating across fields/models with a defined statistical summary.
  2. [§3.2, §2.3, Appendix A] The validation set is not held out in a way that supports the “unseen data” claim. The 8 validation simulations appear to have been used for hyperparameter decisions: Section 2.3 states that the λBC update times are “empirical and depend on the number of epochs to reach the best model,” and Appendix A describes extensive hyperparameter tuning for the compression networks. No separate test split is described. Reported PCC/SSIM averages are therefore optimistically biased and may not reflect performance on truly unseen data. The authors should report metrics on a held-out test set, or state clearly that only 8 simulations were used for both tuning and evaluation and provide appropriate caveats.
  3. [§3.4, Tables 1 and 2] All headline comparisons are point estimates without confidence intervals. Table 1 reports PCC and SSIM as averages over 8 validation simulations; Table 2 provides quartiles only for the CO2-area error, not for MSE or PCC. The observed differences between models — e.g., UNet rollT8 versus UNet++ rollT8 — are small and could easily be within simulation-to-simulation variability. Without error bars, paired differences, or per-simulation results, the qualitative ordering of models is not statistically supported.
  4. [§3.4, §3.5, Conclusions item 3] The conclusion that the grid-size-invariant framework “outperformed the Reduced Order Model approach for unseen data” is not supported by a direct quantitative comparison. Table 1 and Table 2 report only grid-size-invariant models; the ROM results appear only as PCC curves in Figures 6 and 8, and no common numerical table is given for validation PCC at fixed rollout horizons. To make the comparison rigorous, the authors should provide a table of validation PCC (and ideally SSIM/MSE) for both ROMs and grid-size-invariant models at the same timesteps, with error bars.
minor comments (5)
  1. [§1] Typo: “predicion network” should be “prediction network.”
  2. [§3.4] The statement “The Pearson correlation remains above a threshold of 0.75 for the 100-timestep prediction in all fields” is based on point averages and should be accompanied by the spread across simulations or a lower confidence bound.
  3. [§3.2, Eq. (13)] The CO2-area metric depends on the arbitrary threshold Cthreshold=0.5; the sensitivity of the reported conclusions to this value is not discussed.
  4. [§2.2, §3.4] The patch-representativeness limitation is acknowledged in Section 2.2, but the manuscript does not discuss how the uniform spatial sampling of 64x64 patches interacts with the 256x256 channel patterns; a simple diagnostic showing patch-level versus full-domain training metrics would address this.
  5. [Appendix A] The reconstruction MSE values for AE and AAE are reported in the text and repeated in Appendix A; consider presenting them in a single table for clarity.

Circularity Check

0 steps flagged

No circularity found: grid-size invariance rests on the external fully-convolutional-network property; reported metrics compare against external GeoChemFOAM ground truth, not against fitted quantities.

full rationale

No circular derivation chain exists. The grid-size-invariance mechanism is explicitly attributed to the external FCN property that convolutional layers have local connections (Long et al., [39]), not to the authors' prior work. The ROM components (CAE, AAE, UNet, UNet++) are standard literature architectures, and the dataset is externally generated by GeoChemFOAM simulations [47]. Reported metrics (PCC, SSIM, CO2-area error) compare model outputs to these external ground-truth fields; no fitted parameter is substituted for the target. Self-citations such as [14], [48], [52], and [60] are contextual/future-work citations and are not used to derive the central grid-size-invariance claim. The manuscript does state that lambda_BC schedules are empirical (Section 2.3) and reports extensive hyperparameter tuning (Appendix A), and it lacks a separate held-out test set; but this is a statistical validity risk, not a by-construction equivalence, because the validation predictions are not forced by the hyperparameters and the ground truth is not derived from the model. The paper's own caveat that training subsamples 'need to be representative' (Section 2.2) identifies an assumption, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim is empirical and not derived; the main imported assumptions are the representativeness of 64×64 subsamples and the accuracy of the GeoChemFOAM reference data. The free parameters are training hyperparameters, some tuned on the same validation set used for reporting. No new physical entities are introduced.

free parameters (5)
  • λBC boundary-loss multiplier schedule = 0 → 0.5 → 1 at epochs 50/100 (UNet), 15/65 (UNet++)
    Tuned empirically to improve boundary errors; different schedules per architecture complicate fair comparison.
  • Rollout length T = 8
    Chosen by hand for rollout training; no sensitivity study is reported.
  • Cthreshold for CO2 area metric = 0.5 (scaled concentration)
    Chosen as the middle of the scaled concentration range; affects EareaCO2 values, though probably not the main ranking.
  • AAE discriminator/encoder update ratio = 2:1
    Manual tuning to stabilize adversarial training; described in Appendix A.
  • Autoencoder latent compression = 4× reduction per dimension (16:1 memory)
    Architecture choice that defines the latent space in which the ROM predictors operate; reconstruction quality is high, but the choice is not varied.
axioms (5)
  • standard math A fully convolutional network trained on 64×64 inputs can be applied to 256×256 inputs without retraining.
    This is a standard property of convolutional networks with no global pooling/dense layers, stated in Section 2.2.
  • domain assumption The 64×64 training subsamples are representative of full-domain physics (channels, dissolution patterns).
    Stated explicitly in Section 2.2; if false, the grid-size-invariant framework cannot recover large-scale behavior.
  • domain assumption GeoChemFOAM simulations provide accurate ground truth for the surrogate models.
    The surrogates are evaluated against these simulations (Section 3.1); systematic errors in the data would propagate into the comparisons.
  • domain assumption Boundary-condition penalty learned during training carries over to full-domain inference.
    Section 2.3 applies the boundary-loss idea to the prediction domain; the transfer from patch boundaries to true physical boundaries is assumed.
  • domain assumption Grid spacing is the same between training and inference.
    Section 2.2 notes that grid spacing must be unchanged; this is respected in the 64×64/256×256 setup under the same 25 µm spacing.

pith-pipeline@v1.3.0-alltime-deepseek · 19256 in / 11625 out tokens · 107864 ms · 2026-08-02T20:43:42.505280+00:00 · methodology

0 comments
read the original abstract

Modelling rock-fluid interaction requires solving a set of partial differential equations (PDEs) to predict the flow behaviour and the reactions of the fluid with the rock on the interfaces. Conventional high-fidelity numerical models require a high resolution to obtain reliable results, resulting in huge computational expense. This restricts the applicability of these models for multi-query problems, such as uncertainty quantification and optimisation, which require running numerous scenarios. As a cheaper alternative to high-fidelity models, this work develops eight surrogate models for predicting the fluid flow in porous media. Four of these are reduced-order models (ROM) based on one neural network for compression and another for prediction. The other four are single neural networks with the property of grid-size invariance; a term which we use to refer to image-to-image models that are capable of inferring on computational domains that are larger than those used during training. In addition to the novel grid-size-invariant framework for surrogate models, we compare the predictive performance of UNet and UNet++ architectures, and demonstrate that UNet++ outperforms UNet for surrogate models. Furthermore, we show that the grid-size-invariant approach is a reliable way to reduce memory consumption during training, resulting in good correlation between predicted and ground-truth values and outperforming the ROMs analysed. The application analysed is particularly challenging because fluid-induced rock dissolution results in a non-static solid field and, consequently, it cannot be used to help in adjustments of the future prediction.

Figures

Figures reproduced from arXiv: 2602.22188 by Ahmed H. Elsheikh, Aniket C. Joshi, Christopher C. Pain, Claire E. Heaney, Donghu Guo, Hannah P. Menke, Nathalie C. Pinheiro.

Figure 1
Figure 1. Figure 1: Diagrams for the two different options of the prediction neural network compared. In the UNet++ diagram, [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Therefore, the model is not only trained to predict the next timestep, but it should also [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: Rollout training: strategy used to improve multiple timesteps inference in grid-size-invariant framework. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Workflow for training the ROMs. The dashed box indicates the adversarial training extra steps, used only in [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Workflow for inference using one of the reduced-order surrogate models. The compression module produces [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Surrogate model using compression - Autoregressive prediction after 100 timesteps in the carbon storage [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Pearson correlation coefficient (PCC) for autoregressive prediction. Comparison of results obtained using [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Surrogate model using grid-size invariance - Autoregressive prediction after 100 timesteps in the carbon [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Pearson correlation coefficient (PCC) for autoregressive prediction. Comparison of results obtained using the [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison between downsampling methods common in image processing, the autoencoder used and the [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

66 extracted references · 7 canonical work pages

  1. [1]

    Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change, volume In Press

    IPCC.Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change, volume In Press. Cambridge University Press, Cambridge, United Kingdom and New York, NY , USA,

    2021

  2. [2]

    Technical summary

    IPCC. Technical summary. InClimate Change 2022: Mitigation of Climate Change. Contri- bution of Working Group III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press, Cambridge, United Kingdom and New York, NY , USA, 2022. doi: 10.1017/9781009157926.002. URLhttps://www.ipcc.ch/report/ ar6/wg3/

    doi:10.1017/9781009157926.002 2022

  3. [3]

    Map of CO 2 storage projects in europe.https://iogpeurope.org/ resource/map-of-eu-ccus-projects/, 2024

    IOGP Europe. Map of CO 2 storage projects in europe.https://iogpeurope.org/ resource/map-of-eu-ccus-projects/, 2024. Poster (map) published online; accessed on 10 Sept 2025

    2024

  4. [4]

    Bastian E. Rapp. Chapter 29 — Computational fluid dynamics. In Bastian E. Rapp, editor, Microfluidics (Second Edition), Micro and Nano Technologies, pages 653–666. Elsevier, Ox- ford, United Kingdom, second edition edition, 2023. ISBN 978-0-12-824022-9. doi: https:// doi.org/10.1016/B978-0-12-824022-9.00049-8. URLhttps://www.sciencedirect.com/ science/art...

    work page doi:10.1016/b978-0-12-824022-9.00049-8 2023

  5. [5]

    Rao.CFD of the Future: Year 2025 and Beyond, pages 779–795

    Akshai Kumar Runchal and Madhukar M. Rao.CFD of the Future: Year 2025 and Beyond, pages 779–795. Springer Singapore, Singapore, 2020. ISBN 978-981-15-2670-1. doi: 10. 1007/978-981-15-2670-1_22. URLhttps://doi.org/10.1007/978-981-15-2670-1

    work page doi:10.1007/978-981-15-2670-1 2025

  6. [6]

    Model reduction of dynamical systems on nonlinear mani- folds using deep convolutional autoencoders, 2019

    Kookjin Lee and Kevin Carlberg. Model reduction of dynamical systems on nonlinear mani- folds using deep convolutional autoencoders, 2019

    2019

  7. [7]

    Non-intrusive Sur- rogate Modeling for Parametrized Time-dependent PDEs using Convolutional Autoencoders, 2021

    Stefanos Nikolopoulos, Ioannis Kalogeris, and Vissarion Papadopoulos. Non-intrusive Sur- rogate Modeling for Parametrized Time-dependent PDEs using Convolutional Autoencoders, 2021. 22 Figure 9: Comparison between downsampling methods common in image processing, the autoencoder used and the adversarial autoencoder used

    2021

  8. [8]

    Convolutional neural networks: an overview and application in radiology.Insights into Imaging, 9, 2018

    Rikiya Yamashita, Mizuho Nishio, Richard Kinh Gian, and Kaori Togashi. Convolutional neural networks: an overview and application in radiology.Insights into Imaging, 9, 2018. ISSN 1869-4101. doi: 10.1007/s13244-018-0639-9. URLhttps://doi.org/10.1007/ s13244-018-0639-9

    doi:10.1007/s13244-018-0639-9 2018

  9. [9]

    Packt Publishing, Birmingham, UK, 2018

    Iffat Zafar, Giounona Tzanidou, Richard Burton, Nimesh Patel, and Leonardo Araujo.Hands- On Convolutional Neural Networks with TensorFlow: Solve computer vision problems with modeling in TensorFlow and Python. Packt Publishing, Birmingham, UK, 2018. ISBN 1789130336. 23

    2018

  10. [10]

    Gonzalez and Maciej Balajewicz

    Francisco J. Gonzalez and Maciej Balajewicz. Deep convolutional recurrent autoencoders for learning low-dimensional feature dynamics of fluid systems.arXiv preprint, 1808.01346, 2018

    Pith/arXiv arXiv 2018

  11. [11]

    Reduced-order modeling of advection-dominated systems with recurrent neural networks and convolutional autoencoders

    Romit Maulik, Bethany Lusch, and Prasanna Balaprakash. Reduced-order modeling of advection-dominated systems with recurrent neural networks and convolutional autoencoders. Physics of Fluids, 33(3), 2021

    2021

  12. [12]

    Kadeethum, F

    T. Kadeethum, F. Ballarin, Y . Choi, D. O’Malley, H. Yoon, and N. Bouklas. Non-intrusive reduced order modeling of natural convection in porous media using convolutional autoen- coders: Comparison with linear subspace techniques.Advances in Water Resources, 160: 104098, 2022. ISSN 0309-1708. doi: https://doi.org/10.1016/j.advwatres.2021.104098. URL https:...

    arXiv 2022

  13. [13]

    A Comprehensive Deep Learning- Based Approach to Reduced Order Modeling of Nonlinear Time-Dependent Parametrized PDEs.Journal of Scientific Computing, 87, 2021

    Stefania Fresca, Luca Dede’, and Andrea Manzoni. A Comprehensive Deep Learning- Based Approach to Reduced Order Modeling of Nonlinear Time-Dependent Parametrized PDEs.Journal of Scientific Computing, 87, 2021. ISSN 1573-7691. doi: 10.1007/ s10915-021-01462-7. URLhttps://doi.org/10.1007/s10915-021-01462-7

    doi:10.1007/s10915-021-01462-7 2021

  14. [14]

    Heaney, Zef Wolffs, Jón Atli Tómasson, Lyes Kahouadji, Pablo Salinas, André Nicolle, Ionel M

    Claire E. Heaney, Zef Wolffs, Jón Atli Tómasson, Lyes Kahouadji, Pablo Salinas, André Nicolle, Ionel M. Navon, Omar K. Matar, Narakorn Srinil, and Christopher C. Pain. An AI- based non-intrusive reduced-order model for extended domains applied to multiphase flow in pipes.Physics of Fluids, 34(5), May 2022. ISSN 1089-7666. doi: 10.1063/5.0088070. URL http:...

    doi:10.1063/5.0088070 2022

  15. [15]

    Evaluating a deep learning-based surrogate model for predicting wind distribution in urban microclimate design.Building and Environment, 269:112426, 2025

    Houzhi Wang, Wei Ma, Jianlei Niu, and Ruoyu You. Evaluating a deep learning-based surrogate model for predicting wind distribution in urban microclimate design.Building and Environment, 269:112426, 2025. ISSN 0360-1323. doi: https://doi.org/10.1016/j. buildenv.2024.112426. URLhttps://www.sciencedirect.com/science/article/pii/ S036013232401268X

    arXiv 2025

  16. [16]

    Surrogate modeling of fluid dynamics with a multigrid inspired neural network architecture.Machine Learning with Applications, 6:100176, 2021

    Quang Tuyen Le and Chinchun Ooi. Surrogate modeling of fluid dynamics with a multigrid inspired neural network architecture.Machine Learning with Applications, 6:100176, 2021. ISSN 2666-8270. doi: https://doi.org/10.1016/j.mlwa.2021.100176. URLhttps://www. sciencedirect.com/science/article/pii/S2666827021000888

    arXiv 2021

  17. [17]

    A deep-learning super-resolution reconstruction model of turbulent reacting flow.Computers & Fluids, 275: 106249, 2024

    Zhentao Pang, Kai Liu, Hualin Xiao, Tai Jin, Kun Luo, and Jianren Fan. A deep-learning super-resolution reconstruction model of turbulent reacting flow.Computers & Fluids, 275: 106249, 2024. ISSN 0045-7930. doi: https://doi.org/10.1016/j.compfluid.2024.106249. URL https://www.sciencedirect.com/science/article/pii/S0045793024000811

    arXiv 2024

  18. [18]

    Single phase flow simula- tion in porous media by physical-informed unet network based on lattice boltzmann method

    Jiuyu Zhao, Jinsui Wu, Han Wang, Yuxuan Xia, and Jianchao Cai. Single phase flow simula- tion in porous media by physical-informed unet network based on lattice boltzmann method. Journal of Hydrology, 639:131501, 2024. ISSN 0022-1694. doi: https://doi.org/10.1016/j. jhydrol.2024.131501. URLhttps://www.sciencedirect.com/science/article/pii/ S0022169424008977. 24

    arXiv 2024

  19. [19]

    Using U-Net for segmentation of fluid phases in porous media

    Tareq Aljamou. Using U-Net for segmentation of fluid phases in porous media. Master’s thesis, Norwegian University of Science and Technology (NTNU), 2023. URLhttps:// ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/3088500. Accessed: 2025-08-12

    2023

  20. [20]

    Deep residual U-net convolution neural networks with autoregressive strategy for fluid flow predictions in large-scale geosystems.Ad- vances in Water Resources, 150:103878, 2021

    Zhihao Jiang, Pejman Tahmasebi, and Zhiqiang Mao. Deep residual U-net convolution neural networks with autoregressive strategy for fluid flow predictions in large-scale geosystems.Ad- vances in Water Resources, 150:103878, 2021. ISSN 0309-1708. doi: https://doi.org/10.1016/ j.advwatres.2021.103878. URLhttps://www.sciencedirect.com/science/article/ pii/S03...

    arXiv 2021

  21. [21]

    Data- driven methods for flow and transport in porous media: A review.International Jour- nal of Heat and Mass Transfer, 235:126149, 2024

    Guang Yang, Ran Xu, Yusong Tian, Songyuan Guo, Jingyi Wu, and Xu Chu. Data- driven methods for flow and transport in porous media: A review.International Jour- nal of Heat and Mass Transfer, 235:126149, 2024. ISSN 0017-9310. doi: https://doi. org/10.1016/j.ijheatmasstransfer.2024.126149. URLhttps://www.sciencedirect.com/ science/article/pii/S0017931024009797

    arXiv 2024

  22. [22]

    Gege Wen, Meng Tang, and Sally M. Benson. Towards a predictor for co2 plume migration using deep neural networks.International Journal of Greenhouse Gas Control, 105:103223,

  23. [23]

    Unet++: A Nested U-Net Architecture for Medical Image Segmentation

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A Nested U-Net Architecture for Medical Image Segmentation. InDeep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, pages 3–11. Springer, Switzerland, 2018

    2018

  24. [24]

    doi: https://doi.org/10.1016/j.ijggc.2020.103223

    ISSN 1750-5836. doi: https://doi.org/10.1016/j.ijggc.2020.103223. URLhttps: //www.sciencedirect.com/science/article/pii/S1750583620306484

    arXiv 2020

  25. [25]

    Bousias Alexakis and C

    E. Bousias Alexakis and C. Armenakis. Evaluation of unet and unet++ architectures in high resolution image change detection applications.The International Archives of the Photogram- metry, Remote Sensing and Spatial Information Sciences, XLIII-B3-2020:1507–1514, 2020. doi: 10.5194/isprs-archives-XLIII-B3-2020-1507-2020. URLhttps://isprs-archives. copernic...

    work page doi:10.5194/isprs-archives-xliii-b3-2020-1507-2020 2020

  26. [26]

    UNet++: Redesigning Skip Connections to Exploit Multiscale Features in Image Segmenta- tion.IEEE Transactions on Medical Imaging, 2019

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. UNet++: Redesigning Skip Connections to Exploit Multiscale Features in Image Segmenta- tion.IEEE Transactions on Medical Imaging, 2019

    2019

  27. [27]

    Pix2pix++: An en- hanced gans based model for portrait to pencil sketch translation

    Humza Fazal Abbasi, Merium Fazal Abbasi, and Faizan Hamayat. Pix2pix++: An en- hanced gans based model for portrait to pencil sketch translation. In2024 18th International Conference on Open Source Systems and Technologies (ICOSST), pages 1–6, 2024. doi: 10.1109/ICOSST64562.2024.10871136. 25

    arXiv 2024

  28. [28]

    Jimin Park, Yejin Kim, Sang Seob Kim, Kwang Yeom Kim, and Tae Sup Yun. Effect of injection rate and viscosity on stimulated fracture in granite: Extraction of fracture by convo- lutional neural network and morphological analysis.Rock Mechanics and Rock Engineering, 57-3:2159–2174, 2024. doi: 10.1007/s00603-023-03678-5. URLhttps://doi.org/10. 1007/s00603-0...

    work page doi:10.1007/s00603-023-03678-5 2024

  29. [29]

    Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural Ordinary Differential Equations, 2019. URLhttps://arxiv.org/abs/1806.07366

    Pith/arXiv arXiv 2019

  30. [30]

    Ad- versarial autoencoders.arXiv preprint, 1511.05644, 2016

    Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, and Brendan Frey. Ad- versarial autoencoders.arXiv preprint, 1511.05644, 2016

    Pith/arXiv arXiv 2016

  31. [31]

    Raissi, P

    M. Raissi, P. Perdikaris, and G.E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial dif- ferential equations.Journal of Computational Physics, 378:686–707, 2019. ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2018.10.045. URLhttps://www.sciencedirect.com/ scie...

    doi:10.1016/j.jcp.2018.10.045 2019

  32. [32]

    On Neural Differential Equations, 2022

    Patrick Kidger. On Neural Differential Equations, 2022. URLhttps://arxiv.org/abs/ 2202.02435

    Pith/arXiv arXiv 2022

  33. [33]

    Physics-informed neural networks (PINNs) for fluid mechanics: a review.Acta Mechanica Sinica, 37, 2021

    Shengze Cai, Zhiping Mao, Zhicheng Wang, Minglang Yin, and George Em Karniadakis. Physics-informed neural networks (PINNs) for fluid mechanics: a review.Acta Mechanica Sinica, 37, 2021. ISSN 1614-3116. doi: 10.1007/s10409-021-01148-1. URLhttps://doi. org/10.1007/s10409-021-01148-1

    doi:10.1007/s10409-021-01148-1 2021

  34. [34]

    Hesthaven, and Chuhua Zhang

    Wenqian Chen, Qian Wang, Jan S. Hesthaven, and Chuhua Zhang. Physics-informed ma- chine learning for reduced-order modeling of nonlinear problems.Journal of Computational Physics, 446:110666, 2021. ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2021.110666. URLhttps://www.sciencedirect.com/science/article/pii/S0021999121005611

    arXiv 2021

  35. [35]

    Gege Wen, Zongyi Li, Kamyar Azizzadenesheli, Anima Anandkumar, and Sally M. Benson. U-FNO—An enhanced Fourier neural operator-based deep-learning model for multiphase flow.Advances in Water Resources, 163:104180, 2022. ISSN 0309-1708. doi: https://doi.org/ 10.1016/j.advwatres.2022.104180. URLhttps://www.sciencedirect.com/science/ article/pii/S0309170822000562

    arXiv 2022

  36. [36]

    Arthurs and Andrew P

    Christopher J. Arthurs and Andrew P. King. Active training of physics-informed neu- ral networks to aggregate and interpolate parametric solutions to the navier-stokes equa- tions.Journal of Computational Physics, 438:110364, 2021. ISSN 0021-9991. doi: https: //doi.org/10.1016/j.jcp.2021.110364. URLhttps://www.sciencedirect.com/science/ article/pii/S00219...

    arXiv 2021

  37. [37]

    Benchmarking autoregressive conditional dif- fusion models for turbulent flow simulation, 2024

    Georg Kohl, Li-Wei Chen, and Nils Thuerey. Benchmarking autoregressive conditional dif- fusion models for turbulent flow simulation, 2024. URLhttps://arxiv.org/abs/2309. 01745

    2024

  38. [38]

    Nathan Kutz, and Steven L

    Bethany Lusch, J. Nathan Kutz, and Steven L. Brunton. Deep learning for universal linear embeddings of nonlinear dynamics.Nature Communications, 9(1):4950, 2018. ISSN 2041-1723. doi: 10.1038/s41467-018-07210-0. URLhttps://doi.org/10.1038/ s41467-018-07210-0

    doi:10.1038/s41467-018-07210-0 2018

  39. [39]

    Fully convolutional networks for se- mantic segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for se- mantic segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3431–3440, 2015

    2015

  40. [40]

    Teixeira, and Debdipta Goswami

    Indranil Nayak, Ananda Chakrabarti, Mrinal Kumar, Fernando L. Teixeira, and Debdipta Goswami. Temporally-consistent koopman autoencoders for forecasting dynamical systems. 26 Scientific Reports, 15(1):22127, 2025. ISSN 2045-2322. doi: 10.1038/s41598-025-05222-7. URLhttps://doi.org/10.1038/s41598-025-05222-7

    doi:10.1038/s41598-025-05222-7 2025

  41. [41]

    Neural operator: learning maps between function spaces with applications to pdes.J

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: learning maps between function spaces with applications to pdes.J. Mach. Learn. Res., 24(1), January 2023. ISSN 1532-4435

    2023

  42. [42]

    Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhat- tacharya, Andrew M

    Zong-Yi Li, Nikola B. Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhat- tacharya, Andrew M. Stuart, and Anima Anandkumar. Fourier neural operator for para- metric partial differential equations.ArXiv, abs/2010.08895, 2020. URLhttps://api. semanticscholar.org/CorpusID:224705257

    Pith/arXiv arXiv 2010

  43. [43]

    Orhan and Y

    S. Orhan and Y . Bastanlar. Training convolutional neural networks with image patches for object localisation.Electronics Letters, 54(7):424–426, 2018. doi: 10.1049/el.2017.4725. URLhttps://doi.org/10.1049/el.2017.4725

    arXiv 2018

  44. [44]

    Belnet: Basis enhanced learning, a mesh-free neural operator, 2022

    Zecheng Zhang, Wing Tat Leung, and Hayden Schaeffer. Belnet: Basis enhanced learning, a mesh-free neural operator, 2022. URLhttps://arxiv.org/abs/2212.07336

    Pith/arXiv arXiv 2022

  45. [45]

    Patch- cnn: Training data-efficient deep learning for high-fidelity diffusion tensor estimation from minimal diffusion protocols, 2023

    Tobias Goodwin-Allcock, Ting Gong, Robert Gray, Parashkev Nachev, and Hui Zhang. Patch- cnn: Training data-efficient deep learning for high-fidelity diffusion tensor estimation from minimal diffusion protocols, 2023. URLhttps://arxiv.org/abs/2307.01346

    Pith/arXiv arXiv 2023

  46. [46]

    A patch-based convolutional neu- ral network for remote sensing image classification.Neural Networks, 95:19–28, 2017

    Atharva Sharma, Xiuwen Liu, Xiaojun Yang, and Di Shi. A patch-based convolutional neu- ral network for remote sensing image classification.Neural Networks, 95:19–28, 2017. ISSN 0893-6080. doi: https://doi.org/10.1016/j.neunet.2017.07.017. URLhttps://www. sciencedirect.com/science/article/pii/S0893608017301806

    work page doi:10.1016/j.neunet.2017.07.017 2017

  47. [47]

    Julien Maes, Cyprien Soulaine, and Hannah P. Menke. Improved volume-of-solid formula- tions for micro-continuum simulation of mineral dissolution at the pore-scale.Frontiers in Earth Science, 10, 2022. doi: 10.3389/feart.2022.917931

    arXiv 2022

  48. [48]

    Kanatsoulis, and Alejandro Ribeiro

    Damian Owerko, Charilaos I. Kanatsoulis, and Alejandro Ribeiro. Solving large-scale spatial problems with convolutional neural networks, 2024. URLhttps://arxiv.org/abs/2306. 08191

    2024

  49. [49]

    IE-NeRF: Exploring transient mask inpainting to enhance neural radiance fields in the wild.Neurocomputing, 618: 129112, 2025

    Shuaixian Wang, Haoran Xu, Yaokun Li, Jiwei Chen, and Guang Tan. IE-NeRF: Exploring transient mask inpainting to enhance neural radiance fields in the wild.Neurocomputing, 618: 129112, 2025. ISSN 0925-2312. doi: https://doi.org/10.1016/j.neucom.2024.129112. URL https://www.sciencedirect.com/science/article/pii/S0925231224018836. 27

    arXiv 2025

  50. [50]

    Donghu Guo, N. C. Pinheiro, C. E. Heaney, and C. C. Pain. Domain-Agnostic AI Surrogates: Implicit Prediction for Transient Flow with Size and Geometry Invariance.In preparation, 2026

    2026

  51. [51]

    Image-based artificial intelligence empowered surrogate model and shape morpher for real-time blank shape optimisation in the hot stamping process, 2022

    Haosu Zhou and Nan Li. Image-based artificial intelligence empowered surrogate model and shape morpher for real-time blank shape optimisation in the hot stamping process, 2022. URL https://arxiv.org/abs/2212.05885

    Pith/arXiv arXiv 2022

  52. [52]

    Surrogate modeling of melt pool temperature field using deep learning

    AmirPouya Hemmasian, Francis Ogoke, Parand Akbari, Jonathan Malen, Jack Beuth, and Amir Barati Farimani. Surrogate modeling of melt pool temperature field using deep learning. Additive Manufacturing Letters, 5:100123, 2023. ISSN 2772-3690. doi: https://doi.org/10. 1016/j.addlet.2023.100123. URLhttps://www.sciencedirect.com/science/article/ pii/S277236902300004X

    arXiv 2023

  53. [53]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. InProceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2, NIPS’14, page 2672–2680, Cambridge, MA, USA, 2014. MIT Press

    2014

  54. [54]

    Toby R. F. Phillips, Claire E. Heaney, Paul N. Smith, and Christopher C. Pain. An autoencoder-based reduced-order model for eigenvalue problems with application to neutron diffusion.International Journal for Numerical Methods in Engineering, 122(15):3780–3811,

  55. [55]

    URLhttps://onlinelibrary.wiley.com/ doi/abs/10.1002/nme.6681

    doi: https://doi.org/10.1002/nme.6681. URLhttps://onlinelibrary.wiley.com/ doi/abs/10.1002/nme.6681

    work page doi:10.1002/nme.6681

  56. [56]

    Curriculum learning: A survey, 2022

    Petru Soviany, Radu Tudor Ionescu, Paolo Rota, and Nicu Sebe. Curriculum learning: A survey, 2022. URLhttps://arxiv.org/abs/2101.10382

    Pith/arXiv arXiv 2022

  57. [57]

    Generative ad- versarial networks and adversarial autoencoders: Tutorial and survey.arXiv preprint arXiv:2111.13282, 2021

    Benyamin Ghojogh, Ali Ghodsi, Fakhri Karray, and Mark Crowley. Generative ad- versarial networks and adversarial autoencoders: Tutorial and survey.arXiv preprint arXiv:2111.13282, 2021

    Pith/arXiv arXiv 2021

  58. [58]

    Curriculum learn- ing

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learn- ing. InProceedings of the 26th Annual International Conference on Machine Learning, ICML ’09, page 41–48, New York, NY , USA, 2009. Association for Computing Machin- ery. ISBN 9781605585161. doi: 10.1145/1553374.1553380. URLhttps://doi.org/10. 1145/1553374.1553380

    arXiv 2009

  59. [59]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing, 13(4):600– 612, 2004. doi: 10.1109/TIP.2003.819861

    arXiv 2004

  60. [60]

    Menke, and Julien Maes

    Alexandros Patsoukis Dimou, Hannah P. Menke, and Julien Maes. Benchmarking the Viabil- ity of 3D Printed Micromodels for Single Phase Flow Using Particle Image Velocimetry and Direct Numerical Simulations.Transport in Porous Media, 141, 2022. ISSN 1573-1634. doi: 10.1007/s11242-021-01718-8. URLhttps://doi.org/10.1007/s11242-021-01718-8

    work page doi:10.1007/s11242-021-01718-8 2022

  61. [61]

    Boslaugh.Statistics in a nutshell

    Sarah. Boslaugh.Statistics in a nutshell. In a nutshell. O’Reilly, Beijing, second edition. edition, 2012. ISBN 9781449361129

    2012

  62. [62]

    Manning Publications, Shelter Island, New York, 1st edition edition, 2019

    Jakub Langr and Vladimir Bok.GANs in action : deep learning with generative adversarial networks. Manning Publications, Shelter Island, New York, 1st edition edition, 2019. ISBN 9781638354239

    2019

  63. [63]

    Elsheikh

    Marcos Cirne, Hannah Menke, Alhasan Abdellatif, Julien Maes, Florian Doster, and Ahmed H. Elsheikh. A deep-learning iterative stacked approach for prediction of reactive dissolution in porous media, 2025. URLhttps://arxiv.org/abs/2503.08410. 28

    arXiv 2025

  64. [64]

    Unsupervised representation learning with deep convolutional generative adversarial networks, 2016

    Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks, 2016. URLhttps://arxiv.org/abs/ 1511.06434

    Pith/arXiv arXiv 2016

  65. [66]

    Burt and E

    P. Burt and E. Adelson. The Laplacian Pyramid as a Compact Image Code.IEEE Transactions on Communications, 31(4):532–540, 1983. doi: 10.1109/TCOM.1983.1095851. 29

    arXiv 1983

  66. [2021]

    URLhttps://www.ipcc.ch/report/ar6/wg1/

    doi: 10.1017/9781009157896. URLhttps://www.ipcc.ch/report/ar6/wg1/

    doi:10.1017/9781009157896